Enterprise agent platform for building, deploying, and operating AI teammates.
"""Shared embedding stubs for the LanceDB memory-store tests (#822).
``ConstantEmbedding`` returns the same vector for every text, so vector distance
cannot rank results — isolating where-prefilter / top-k selection from similarity
ranking (used by the population-independence and column-promotion suites).
``VaryingEmbedding`` maps each registered text to a distinct vector, so cosine/L2
distance is meaningful and similarity ordering can be asserted *alongside* the
where-prefilter (used by the ranking-under-prefilter test).
"""
from __future__ import annotations
from typing import Any, Mapping, Sequence
from xagent.core.model.embedding import BaseEmbedding
class ConstantEmbedding(BaseEmbedding):
"""Every text embeds to the same vector (configurable dimension/value)."""
def __init__(self, dim: int = 8, value: float = 0.1) -> None:
self._dimension = dim
self._value = value
def encode(self, text: Any, dimension: Any = None, instruct: Any = None) -> Any:
if isinstance(text, str):
return [self._value] * self._dimension
return [[self._value] * self._dimension for _ in text]
def get_dimension(self) -> int:
return self._dimension
@property
def abilities(self) -> list[str]:
return ["embed"]
class VaryingEmbedding(BaseEmbedding):
"""Maps each registered text to a distinct vector so distance ranks results;
an unregistered text falls back to a fixed zero vector."""
def __init__(self, vectors: Mapping[str, Sequence[float]], dim: int) -> None:
self._vectors = {key: list(value) for key, value in vectors.items()}
self._dimension = dim
self._default = [0.0] * dim
def _one(self, text: str) -> list[float]:
return list(self._vectors.get(text, self._default))
def encode(self, text: Any, dimension: Any = None, instruct: Any = None) -> Any:
if isinstance(text, str):
return self._one(text)
return [self._one(item) for item in text]
def get_dimension(self) -> int:
return self._dimension
@property
def abilities(self) -> list[str]:
return ["embed"]